Skip to content

Deep Learning

Under development

This lesson is part of the course scaffold and is being actively written. The learning objectives and outline below define its final scope.

Learning objectives

By the end of this lesson you will be able to:

  • Implement feed-forward networks, LSTMs, temporal convolutional networks, and transformers for financial sequence problems
  • Diagnose why deep models underperform gradient boosting on daily-bar problems in terms of sample size, signal-to-noise, and non-stationarity
  • Identify the settings — intraday data, cross-sectional scale, alternative data — where deep learning plausibly earns its keep
  • Benchmark every deep model against the tree-ensemble and linear baselines from the previous lesson and report the comparison honestly

Outline

  1. Feed-forward networks on tabular features — the honest starting point
  2. LSTMs — sequence modeling and its cost in data
  3. Temporal convolutional networks
  4. Transformers for financial sequences — attention, positional encoding, and hype control
  5. Why deep learning underwhelms on daily bars — samples, noise, and regime shift
  6. Where deep learning earns its keep — intraday, cross-sectional, and alternative data
  7. Benchmarking discipline — deep models vs the baselines that came before

Prerequisites